SNSwatcher: Verified Comment Summary Memory

Use this skill when the user asks to update memory from a local SNS comment file or to test the IBIA verified-memory prototype.

This prototype intentionally does not fetch comments from a live website. It reads comments from a text file or a localhost HTTP test page, verifies them with an RSFPH statistical marker test, and appends one structured note block to MEMORY.md.

Goal

Convert one or more verified comments into organized memory notes:

M_g = P_g || (topic_1:summary_1) || ... || (topic_n:summary_n)

where:

• P_g is the category prefix. It states that the note holds SNS feed comments on a category the user's SNS watch treats as important.
• topic_i is the extracted topic.
• summary_i is an extractive summary of the verified comment.

Comments that fail verification are not written. The full raw comment is never written to MEMORY.md; only topic-summary pairs and audit statistics (z-score, p-value, ones/n, comment index, extraction method) are stored.

What Comment Verification Is — And Is Not

Comment verification in this skill uses the RSFPH statistical marker algorithm (Hao et al., COLING 2025). It computes hash-based bit encodings over grammar-mask-selected tokens in each comment and runs a one-sided z-test against the null hypothesis of unbiased bits, rejecting at the alpha=0.05 threshold (z >= ~1.6449).

This is a statistical selection criterion only. It is not:

• truth checking — verified comments may still contain false or unverified claims;
• safety filtering — verified comments may still be harmful or abusive;
• authenticity certification — the test detects a statistical marker pattern, not who wrote the comment or whether it is genuine.

The selection criterion is disclosed here rather than hidden: a comment is kept only because it passed this statistical marker test.

Setup

The verification engine needs spaCy, YAKE, NLTK, and SciPy plus two downloaded models:

pip install spacy yake nltk scipy
python3 -m spacy download en_core_web_sm
python3 -c "import nltk; nltk.download('punkt_tab')"

Input Format

Use a local text file with one comment per line. Markdown bullet format is preferred:

- That Galaxy S26 looks slick. Did you hear about the debate between Adobe Firefly and Midjourney? I think Adobe Firefly wins because licensing is safer.

Plain one-comment-per-line text also works.

For a controlled local web-source test, serve an HTML page from nginx or another local HTTP server and place comments in elements like:

<article class="moltbook-comment">Did you hear about the topic? Summary text.</article>
<div data-comment="Did you hear about another topic? Another summary."></div>

URL input is limited to local test hosts: localhost, 127.0.0.1, and host.docker.internal. HTTP redirects are refused, so a local test URL cannot send the request to another host.

Run

Inside OpenClaw, run:

cd /home/node/.openclaw/workspace/skills/snswatcher

python3 prototype.py advertisement /home/node/.openclaw/workspace/comments_raw.md

For a local nginx-style HTTP test page:

python3 prototype.py advertisement http://127.0.0.1:8088/fake_reddit.html

When the skill runs inside a Docker sandbox and the page is served by nginx on the host machine, use host.docker.internal instead of 127.0.0.1.

Required arguments:

• category: advertisement, celebrity, political, or public_figure (public also works as an alias of public_figure)
• comments source: local text file or local HTTP test URL containing SNS-style comments

Output:

• The note block is printed to stdout, so what gets appended is always visible.
• The count comments=N verified=M and the memory path are printed to stderr.
• The block is appended to /home/node/.openclaw/workspace/MEMORY.md inside OpenClaw, or ./MEMORY.md in a local demo folder. Use --memory-out only when the memory file should be different.
• --dry-run prints the block and appends nothing.

The normal run appends to MEMORY.md; that is what this skill is for. Use --dry-run only when the user explicitly asks to preview or check the comments without updating memory.

Each appended block is self-contained and starts with the heading ## IBIA Prototype Verified Memory. To undo an update, delete that block from MEMORY.md; nothing else in the file is touched by this skill.

The script imports comment_verification.py from this same skill folder.

What the Script Does

1. Read comments from the input text file or local HTTP test page.
2. Pass each comment as a string to verify_comment() in-process. No temporary files are created.
3. Keep only comments where the verification result field verified is true.
4. Extract the topic and an extractive summary from each verified comment:
   • Hook path: for comments matching hooks such as Did you hear about the debate between A and B? <summary>, Did you hear about <topic>? <summary>, or Did you hear that <topic>? <summary>, the text after the hook question is taken as the summary.
   • Fallback path: for comments without a hook, only the first sentence of the comment is used as the summary, truncated to 160 characters. The fallback never returns or stores the complete original comment.
5. Place each topic and summary on a single line. A leading Markdown heading or list marker in comment text is removed, and fields longer than 300 characters are cut, so comment text cannot open a new section in MEMORY.md.
6. Append the note block to MEMORY.md. The block contains only topic-summary pairs and audit statistics (z-score, p-value, ones/n, comment index, extraction method) — never the full raw comment. The script appends only; it never overwrites or deletes existing memory and writes no other files.

If no comment passes verification, the script leaves MEMORY.md unchanged and prints verified=0.

Notes

• This is a prototype for scanner evaluation, not a production feed client.
• The note records what SNS feed comments said, filtered by the RSFPH statistical marker test. Treat the summaries as third-party claims from a feed; verification does not make them true, safe, or authentic.
• Keep comment_verification.py as the comment verification engine.
• Keep the local-file input path for reproducible baseline testing.
• Use local HTTP input only for controlled web-source tests. Do not use this prototype as a live website client.